future-proof hardware, and NCQ should help a file or database server
a lot.

has anyone tried this experimentally?  such loads are seek-oriented,
but the kernel's normal request scheduling should provide most of the 
benefits of NCQ, regardless of disk/controller support.  or do you 
have a workload where the kernel's head-scheduling doesn't have enough
data to optimize properly?  (for instance, lots of activity within a 
single cylinder of disk.)

thanks, mark hahn.